AI already changes cosmetic formulation in three narrow places: drafting a starting formula, searching ingredient interactions, and screening candidate molecules. It does not replace a formulator. In the ChemBench evaluation of over 2,700 chemistry questions, leading language models beat expert chemists on average yet failed basic tasks and stayed overconfident.
- 2,700+ question–answer pairs in the ChemBench benchmark (peer-reviewed study, Nature Chemistry, 2025): the best models outscored the best human chemists on average, while still failing simple tasks and reporting high confidence when wrong.
- Below 1 % — no order, no numbers. Under Article 19 of Regulation 1223/2009 as retained in UK law, ingredients present at under 1 % may be listed in any order after those above 1 %, and concentrations are never declared on pack. An AI reading an INCI list cannot recover a formula from it.
- Homosalate 10 %, Benzophenone-3 2.2 % / 6 % — the limits currently in Annex VI of the UK Cosmetics Regulation (regulation text). The EU figure for homosalate diverged after a later amendment, which is exactly the class of detail a language model states confidently and gets wrong.
- 1.0–2.0 % as supplied (w/w) — the recommended use level for INOLEX SpectraStat GHL Natural, narrowing to 1.0–1.5 % in an O/W emulsion, added to the water phase below 80 °C (supplier technical page). A number tied to one raw material and one system, not a universal rule.
- 6–20 % as supplied, pH below 5 — the conditioning system INOLEX AminoSensyl HC in rinse-off and leave-in conditioners, added to the oil phase at 75 °C (supplier technical page). Change the pH and the supplier's own range stops applying.
Imagine this: you open your laptop at 11 PM, you don't have access to your supplier's database, and the deadline for a new formulation is tomorrow morning. In the past, this meant hours buried in stacks of textbooks or desperate searches through forums. Today, some cosmetic chemists simply open ChatGPT—and get a starting formula in three minutes. This isn't magic, and it isn't a threat to the profession. It is a tool that is already reshaping the industry—and the sooner you understand how it works (and where it lies), the better it will be for your formulas.
What AI can do in cosmetic chemistry right now
Talk of artificial intelligence in the beauty industry often boils down to personalized recommendations on retailer websites. But that is just the tip of the iceberg. At the level of formula development, AI is already performing several specific functions—and some of them are truly impressive.
It is worth separating two very different kinds of evidence before going further, because most confusion about "AI in cosmetics" comes from mixing them. A peer-reviewed study tells you what a model did on a defined task with a defined scoring rule. A supplier technical page tells you what that supplier's own raw material does in that supplier's own test system. A vendor platform announcement tells you what a company would like you to believe about software it sells, and is a marketing claim until someone publishes the method. This article labels which of the three is speaking every time a number appears.
Generating starting formulations
Large language models—GPT-4, Claude, Gemini—are trained on massive amounts of text, including patents, scientific articles, ingredient technical data sheets, and supplier formularies. This means that if you ask the model to suggest a basic recipe for a moisturizing shampoo, it will output something like: Sodium Laureth Sulfate 10–12%, Cocamidopropyl Betaine 3–5%, Sodium Chloride for viscosity adjustment, panthenol 0.5%, fragrance, preservative—and this will be a reasonable starting point. Not perfect, but a functional base from which you can move forward.
How reasonable, exactly? The best public measurement to date is ChemBench, a peer-reviewed benchmark of more than 2,700 curated chemistry question–answer pairs, in which the strongest models outperformed the best human chemists in the study on average — and simultaneously failed on some basic tasks while returning overconfident predictions (Mirza et al., Nature Chemistry, 2025). Read that result carefully, because it cuts both ways: it is evidence that a model's chemistry is genuinely useful as a draft, and evidence that its confidence carries no information about whether this particular answer is right. The failure mode is not "the model is dumb"; it is "the model is equally fluent when correct and when wrong."
Specialized AI tools for cosmetics—such as Mindsync or GPT-based platforms with cosmetic datasets—go further: they claim to account for ingredient compatibility, pH ranges, and even regional regulatory restrictions. Treat that as a vendor claim rather than a published result: at the time of writing, none of these commercial cosmetic assistants has released a benchmark, a test set, or an error rate that an outside formulator could reproduce. To understand why pH is so critical in any formula, I recommend reading our guide to pH in cosmetics—without this foundation, even the smartest neural network won't save you from an unstable formula.
Searching for ingredients and interactions
One of the most underrated use cases for AI is a quick search for ingredient interactions. Are you adding a new cationic conditioning agent to a formula and want to ensure it doesn't conflict with anionic thickeners? Previously, this required reading five technical data sheets. Now, it takes one prompt in GPT with a specification: "explain the interaction between Polyquaternium-10 and Carbomer in a shampoo with a pH of 5.5."
Of course, the answer must be verified. But as a starting point for research, it saves an hour of work. Especially when it comes to complex systems: for example, how gelling agents and thickeners behave under stress—more on this in our article about tribology, gums, and gelling agents.
Where AI genuinely earns its place beyond chat is in screening — narrowing a large candidate space before anyone touches a beaker. A published example outside the chat window: two ligand-based machine learning models were used to screen natural-product and drug libraries for tyrosinase inhibitors, the top hits were re-ranked by molecular docking, and the three leading candidates were then confirmed in vitro, all showing stronger mushroom-tyrosinase inhibition than arbutin (peer-reviewed study, ACS Omega, 2025). Note what makes that result credible and a chatbot answer not: the model produced a ranked shortlist, and wet-lab assays — not the model — decided which entries survived.
Where AI frankly still falls short
This is where we start an honest conversation. AI is a probabilistic tool. It doesn't "know" chemistry in the way an experienced formulator does. It predicts the most likely next token in a text, relying on patterns in its training data. And this creates several systemic problems.
Hallucinations and outdated data
Models regularly "hallucinate"—they provide non-existent ingredients, incorrect concentrations, or outdated regulatory data with complete confidence. UV filters are the textbook case, because the numbers move and the models do not move with them. In Annex VI of Regulation 1223/2009 as retained in UK law, homosalate currently stands at 10 %, while benzophenone-3 is split by product type at 2.2 % for general skin products and 6 % for face, hand and lip products. The EU version of the same annex has since moved homosalate to a lower, face-only figure. Two jurisdictions, one ingredient, two legal numbers — and a language model asked "what is the maximum permitted level of homosalate" will typically answer with one number, no jurisdiction, and total confidence. That is no small matter: it is a potential product recall.
There is a second, quieter version of the same problem that trips up anyone using AI to reverse-engineer a competitor. A model handed an INCI list will happily "estimate" the percentages. It cannot. Under Article 19 of the same regulation, the list runs in descending order of weight only down to the 1 % line — below that, ingredients may be listed in any order — and concentrations are never printed at all. So everything the label tells you about a preservative at 0.8 % or an active at 0.3 % is the model's invention, not a reading of the pack. This is a regulatory limit on the information available, not a limit that better prompting can lift.
The second painful point is the texture and aesthetics of a cream formula. AI doesn't feel how a cream sits on the skin. It doesn't know that your clients hate a "tacky" finish or that a specific batch of shea butter from Burkina Faso behaves differently than West African shea—even though climate really does affect the composition of plant oils, and that is a nuance that lives in the hands of the formulator, not in a dataset.
Lack of sensory expertise
Cosmetics are, to a huge extent, a sensory experience. Shampoo foam, the slip of a conditioner on wet hair, the feeling after rinsing—all of this is impossible to fully digitize into a training dataset. AI might suggest adding 1% Behentrimonium Chloride to improve detangling—and that is technically correct. But it won't tell you that in your specific system with a high cetyl alcohol content, this will result in an undesirable "waxy" tactile feel. That knowledge comes only from practice.
Compare that vague "1 %" with how a supplier actually states a dose. The conditioning system INOLEX AminoSensyl HC (Brassica Alcohol and Brassicyl Valinate Esylate) is documented on the manufacturer's own technical page at 6–20 % as supplied for rinse-off and leave-in conditioners — with the material added to the oil phase at 75 °C and a requirement that the finished formula sits below pH 5. Every part of that sentence is load-bearing. It is a range for that raw material, in that phase, at that pH, from the company that sells it; it says nothing about a different quat in a different base. A model that averages a dozen such ranges into one "typical 1–5 %" has destroyed precisely the information that made them usable.
Personalization: where AI is truly changing the game
The most revolutionary application of AI in cosmetics is not generating cream formulas, but personalization on an industrial scale. This is something that was previously physically impossible.
Consumer data analysis
Companies like Prose or Function of Beauty use machine learning algorithms to analyse hundreds of parameters: hair porosity, water hardness in the region of residence, scalp type, color treatment, and climate. The result is a personalized formulation that is technically different from your neighbor's. The mechanism is real — it is backed by genuine variation in protein levels (for example, Hydrolyzed Keratin from 0.5% to 3%), types of silicones (Amodimethicone vs Dimethicone), and the ratio of moisturizing agents. The performance, however, is a company claim: neither brand has published the model, the input features, or an efficacy comparison against a non-personalized control, so "hundreds of parameters" is advertising copy until it is measured.
What personalization at scale looks like when someone does publish it is instructive. In a study of 1,156,703 adult Chinese women, an algorithm graded acne severity from high-resolution smartphone selfies alongside a lifestyle questionnaire, and the AI-scored features — blackheads, pore severity, dark circles, skin roughness — tracked acne severity, which fell after age 25, bottomed out at 40–44 and rose again thereafter (peer-reviewed study, Skin Research and Technology, 2024). That is the honest shape of the technology: excellent at finding population-level patterns in a million faces, silent on what any individual should put on hers.
For a home-based formulator or a small brand, this level is currently unattainable, but it is important to understand where the industry is heading right now. Especially if you are just starting your journey — take a look at the about our school page, where you can get an idea of which skills will be in demand in the coming years.
Stability Prediction
Several large manufacturers — Evonik, Givaudan in fragrance, Unilever with internal developments — describe ML models used to predict emulsion behaviour before real-world testing begins. The principle is straightforward: an algorithm trained on thousands of historical stability tests can rank which combinations are most likely to separate after 6 months at 40 °C, and the failures get filtered out at the modelling stage instead of the bench.
Two editorial cautions, both deliberate. First, these are corporate announcements, not published methods — an internal platform name that appears in a press release and nowhere in the literature should be treated as branding, and this article no longer repeats one that we could not trace to any Evonik document or publication. Second, even where the modelling is real it is a triage step, not a result: nothing in it replaces a real stability protocol at 40 °C, at −10 °C, and through freeze–thaw cycling. Where AI does have a documented, checkable role in cosmetic safety work is narrower and less glamorous — for instance an explainable deep-learning framework built to flag allergenic motifs in peptide sequences before those peptides enter a formulation, with the model's reasoning exposed via SHAP and LIME so a human can audit why a sequence was flagged (peer-reviewed study, Journal of Proteome Research, 2026).
Practical Guide: How to Use AI in Your Formulation Work Right Now
If you formulate — professionally or as a hobby — here are specific scenarios where AI really helps, and where caution is needed.
What You Should Entrust to a Neural Network
- Generating a starting formula. Ask ChatGPT or Claude to suggest a base — and use it as a draft, not as a final recipe. Specify: product type, target pH, desired viscosity, and ingredient restrictions.
- Explaining mechanisms of action. "Explain how Polyquaternium-7 adsorbs onto the surface of damaged hair" — this is an excellent prompt. The model will provide a clear explanation that you can verify against primary sources.
- Searching for INCI synonyms. When you need to find trade names for a specific INCI name or vice versa — AI handles this quickly. Confirm the result against the nomenclature held by the Personal Care Products Council, which owns the INCI naming system, rather than against the model's memory.
- Brainstorming alternatives. "What can replace DMDM Hydantoin in this system while maintaining a broad spectrum of activity?" — this is a good question for generating ideas that you must then verify. Verification means going to the actual document: INOLEX SpectraStat GHL Natural (Propanediol, Caprylhydroxamic Acid, Glyceryl Heptanoate), for example, is documented by its manufacturer at 1.0–2.0 % as supplied (w/w), narrowing to 1.0–1.5 % in an O/W emulsion and 1.5–2.0 % in cleansers and sunscreens, added to the water phase below 80 °C, with pass results quoted against EP-A, EP-B, USP 51 and PCPC challenge tests. That is a supplier statement about a supplier's raw material — a real number with a stated test method, which is exactly what an AI-suggested "use around 1 %" is missing.
- Help with product descriptions. Copywriting for cosmetics is another strong suit of language models.
Where Human Expertise Is Required
- Final regulatory compliance check. Always check the current text directly: the UK Cosmetics Regulation (Regulation 1223/2009 as retained in UK law) for products placed on the GB market, the EU version of the same regulation together with the CosIng database where you also sell into the EU, and the IFRA Standards for anything carrying fragrance. Never accept a limit from a model without opening the annex.
- Evaluation of texture and sensory properties. This requires hands, a nose, and skin — no algorithm can replace that.
- Stability testing. Real tests at 40°C, -10°C, and under freeze-thaw cycling conditions are mandatory. AI predictions are only an auxiliary tool.
- Working with new or rare ingredients. If an ingredient has appeared recently, it is almost non-existent in the model's training data — here, you need technical data sheets and direct contact with the supplier.
Ethics and the future: AI as a co-author, not a replacement
One of the most common fears I hear from aspiring cosmetic chemists is: "Why learn to formulate if AI will do it for me?" That is roughly like asking in the year 2000: "Why learn photography if there are digital cameras?" Technology changes the tools, but it does not eliminate the need to understand what you are doing.
The ChemBench result gives that answer an evidence base rather than a sentiment. If a model can beat expert chemists on average and still fail elementary questions with full confidence, then the scarce skill is no longer recall — it is the ability to tell which of two equally fluent answers is the broken one. That is a chemistry skill, and it is acquired at the bench.
AI does not bear responsibility for product safety. It does not know that a specific batch of jojoba oil from your supplier has a non-standard fatty acid profile. It does not feel that a cream formula is "almost right," but that something is off — that intuitive knowledge that comes after hundreds of experiments. That is precisely why the path from curiosity to professional formulation still requires real training — we wrote about this in the article how to become a cosmetic chemist.
The future that is already arriving is a hybrid model. AI takes on the routine: searching, generating options, documentation, and preliminary compatibility analysis. The human takes on judgment, sensory evaluation, ethics, and the final decision. Those who master both languages — chemistry and algorithms — will be in the strongest position.
If you want to build exactly that foundation — to understand chemistry well enough to be able to critically evaluate what any tool, including a neural network, suggests — take a look at the Walker Formulation Academy Club. It is a community of practitioners who analyse real cases, not theory in a vacuum.
Can you fully trust ChatGPT when developing a shampoo or hair conditioner formula?
No — and this is not a flaw of the tool, but its nature. ChatGPT generates probabilistic responses based on patterns in its training data, and the ChemBench study found that the strongest models stay overconfident even where they fail. The starting formula it suggests may be a reasonable starting point, but it requires verification across several parameters: the relevance of regulatory restrictions, the compatibility of specific ingredient brands (not just INCI), and stability in your production system. Treat an AI-generated formula like a draft from an intern with a good theoretical background — it is useful, but it requires expert verification.
Which AI tools are cosmetic chemists actually using in 2024–2025?
Among publicly available tools, general-purpose assistants such as ChatGPT and Claude are used for generating formulas, explaining mechanisms, and searching for ingredients. Among specialized tools, there are RAG-based (Retrieval-Augmented Generation) platforms trained on cosmetic datasets, and several raw material suppliers have begun integrating AI assistants into their portals for technologists — capabilities announced by the companies themselves rather than measured in the literature. For stability prediction and molecular modeling, more specialized computational chemistry software is used, but this is at the level of large R&D laboratories. For home crafters and small brands, a general model plus critical thinking is currently sufficient.
Will AI change the educational requirements for a cosmetic chemist?
It is more likely to shift the focus rather than eliminate the need for education. A basic understanding of emulsion chemistry, surfactants, pH balance, and stability will remain fundamental—precisely because it is impossible to evaluate the quality of what the algorithm suggests without it. However, the value of data skills, the ability to formulate precise queries (prompt engineering), and the critical evaluation of AI outputs will increase. A chemist who possesses both skill sets will be significantly more productive.
Why can't an AI work out a competitor's percentages from the INCI list?
Because the information is not on the label. Article 19 of Regulation 1223/2009 requires ingredients to be listed in descending order of weight only above 1 %; everything below 1 % may appear in any order, and no concentration figure is declared anywhere on pack. Any percentage a model gives you for a competitor product is generated, not read.
Is a supplier technical data sheet a better source than a research paper?
It is a different source, not a better one. A technical page such as INOLEX's for SpectraStat GHL Natural documents what that specific raw material did in that supplier's own challenge tests, at a stated use level and process temperature — authoritative for that material, and silent about anyone else's. A peer-reviewed study describes a defined population and endpoint. Use the technical page for dosing and processing, the study for whether the effect exists at all.
Sources
- Mirza A. et al. A framework for evaluating the chemical knowledge and reasoning abilities of large language models against the expertise of chemists. Nature Chemistry, 2025 (PMID 40394186)
- Regulation (EC) No 1223/2009 on cosmetic products, Article 19 (Labelling), as retained in UK law — legislation.gov.uk
- Regulation (EC) No 1223/2009, Annex VI — UV filters allowed in cosmetic products, as retained in UK law — legislation.gov.uk
- Discovery of Potential Tyrosinase Inhibitors via Machine Learning and Molecular Docking with Experimental Validation. ACS Omega, 2025 (PMID 40918329)
- Facial adult female acne in China: an analysis based on artificial intelligence over one million. Skin Research and Technology, 2024 (PMID 38572573)
- Deciphering Allergen Peptides for Dermatological and Cosmetic Applications with Explainable Artificial Intelligence. Journal of Proteome Research, 2026 (PMID 42345568)
- INOLEX, SpectraStat GHL Natural — product technical page (recommended use levels, process conditions, challenge-test results)
- INOLEX, AminoSensyl HC — product technical page (use level, phase, pH requirement)
